This matches something we keep running into building our own agent skills, that all-or-nothing pattern where a variant either fires on every run or none, no middle ground. What I haven't seen tested here, and it's maybe out of scope for 38 runs, is whether stacking several distinct trigger phrases into one description ever costs you anything on the narrower ones, versus being purely additive. We write fairly dense multi-clause descriptions to cover several ways a request might come in and I've always assumed that's free, but your commit-message-name result, where the name only caught the literal-match request and missed the paraphrase, makes me want to go check whether a long description can dilute the match on a phrasing it's technically supposed to also cover. The cost and turn-count numbers are the part I'll actually use though. Knowing a fired skill roughly doubles turns and cost is a real input when deciding how aggressively to write a description, not just whether it triggers at all.